Cut the AI roadmap to one workflow. Shipped before Q3.
A board mandate to “AI-enable the practice this year” and three vendor proposals waiting on which one to start with.
A regional accounting and advisory firm, twelve partners, around fifty staff. The board had voted at the annual offsite to “AI-enable the practice this calendar year” with a budget line of roughly a hundred and twenty thousand. Three vendors were in the funnel, a workflow-automation platform aimed at audit and assurance, a generative-AI tax-research suite, and a Copilot-style assistant from the firm's existing document-management vendor. Each proposal individually was reasonable. Stacked, they totalled the budget plus a chunk of partner attention nobody had counted. The managing partner asked us to recommend which of the three to start with. The question he was actually asking, less visibly, was whether the mandate itself was sound. The partners were split: half believed the firm was already behind, half believed the noise was overblown. Neither group could name the work that the right tool would help with on a Tuesday morning. They asked us to do the naming.
Don't buy all three. Don't kill the mandate. There is exactly one workflow where the AI win is real this year, pilot that one tool against that one workflow, and let the curve do the rest.
We opened the second hour with the conclusion. The mandate wasn't wrong, there was real AI leverage available to the firm this year, but the three-vendor portfolio approach was. Two of the three products were selling platforms whose value would decline as the foundation models they sat on commoditised through 2026; the third was selling a focused tool against a specific high-leverage workflow that the firm could pilot small and measure clearly. The right move was not a portfolio bet at a hundred and twenty thousand, and not a full pause for a year, both were too large. The right move was a single pilot, sized small, with named success metrics, that would either earn the broader budget next year or honestly say it hadn't. We named the workflow, partner-level audit document review, and named why that pilot worked while the other two products' use cases did not.
We sized the pilot in front of the partner group so the math was theirs, not ours.
We walked the partners through the three proposals against one question: what is each vendor actually selling beyond the underlying model. The audit-automation platform was selling workflow, a UI on top of GPT-class inference, applied to a specific decision the firm makes thousands of times a year (does this clause in this document warrant partner attention or not). That UI plus the workflow logic was real work, real value, and difficult to build in-house at the firm's size. The tax-research suite was selling a curated knowledge base; the firm already had a quarter of that knowledge internally and the other three-quarters was available from the model directly with the same accuracy caveat. The document-management Copilot was selling integration, and that integration would be native, free, and substantially better when the parent vendor shipped it in Q3 anyway. So one of the three was selling a thing the firm could not do itself this year; the other two were not. We sized the pilot. Eighteen thousand for the audit-review tool's annual licence. Fourteen thousand of our advisory, for the change-management piece (training, metric instrumentation, success criteria, and one cycle of writing the success memo to the board). Total under thirty-five thousand. Ninety-day shipping target. Compared to the hundred and twenty thousand the firm was about to commit and what it would have looked like by Q3, the math drew itself.
Declined two vendors. Cut the budget by seventy percent. Named the success metrics the pilot would have to hit to earn next year's broader spend.
We wrote the recommendation to the full partner group, three pages, with the curve on page two and the pilot success metrics on page three. The recommendation declined two of the three vendor proposals outright. It named the third vendor and the pilot scope; it named the budget envelope (a third of the original); it named the success metrics (partner hours saved per engagement, error catch rate against the existing manual review, partner satisfaction at end-of-pilot); it named what would have to be true for the firm to expand the engagement next year and what would have to be true for the firm to declare the pilot inconclusive and shelve the broader mandate. It declined the obvious follow-on for us, a retained “AI advisor” relationship through the rest of the year. The work after the pilot belonged to the senior associate they had identified as the AI champion, not to us. The board approved the recommendation in two weeks. The pilot shipped on day eighty-one of ninety. The partner-hours-saved metric came in above target by the end of Q3. The firm preserved roughly eighty-five thousand of the original budget; that money funded the AI champion role into next fiscal year, when the curve had cleared enough for the broader play to make sense. The two declined vendors, when called and thanked, individually told us they understood. One sent a referral our way the next quarter.
A named person walked the cut to the partner table and said “buy one third of what the board approved, pilot the one workflow, and let next year's spend earn its way.” Not a faceless “we”, not a hedged recommendation that left the door open to all three vendors being phased in. The signed cut was the product. The partner who had been most sceptical of the AI mandate sent the kind of follow-up that exists at firms whose seniors have been waiting, often for years, to be told the loud thing and the right thing were not the same thing.